the open applied informatics journal, 1-10 1 ...the open applied informatics journal, 2007, 1, 1-10...

The Open Applied Informatics Journal, 2007, 1, 1-10 1

1874-1363/07 2007 Bentham Science Publishers Ltd.

Bioinformatics Study of Functional Associations Observed in Multiple Sources of Human Genome Data

Seungwoo Hwang1,2

and Igor B. Kuznetsov*,1

1Gen*NY*sis Center for Excellence in Cancer Genomics, Department of Epidemiology and Biostatistics, University at

Albany, SUNY, One Discovery Drive, Rensselaer, NY 12144, USA

2Korean BioInformation Center, KRIBB, Daejeon 305-806, Korea

Abstract: High-throughput genome analysis techniques produce the ever increasing number of heterogeneous large-scale

datasets. Studies of these mutually complementary sources of data promise insights into a global picture of the living cell.

Here, we present a simple bioinformatics methodology for the analysis of multiple heterogeneous sources of ‘omic’ (ge-

nomic, proteomic, etc) data. We apply this methodology to study associations among four types of human ‘omic’ data:

protein-protein interactions, gene expression, transcription factor binding sites, and functional pathways. The results of

our study indicate that the proposed approach can be used to identify and rank statistically significant functional associa-

tions among genes. We show that combinations of multiple data types provide additional insights into the properties of

functional pathways. The proposed methodology can also be used as a quantitative procedure for evaluating the quality of

‘omic’ datasets.

INTRODUCTION

Recent technological advances in high-throughput data acquisition provide us with various types of large-scale datasets, such as whole genome sequences, gene expression, protein-protein interactions, functional pathways, location of transcription factor binding sites, etc. This gives us a unique opportunity to use bioinformatics to integrate diverse and mutually complementary sources of ‘omic’ data (e.g., ge-nomic, proteomic, etc.) into a single coherent systems biol-ogy framework in order to provide functional inference, re-veal essential features of gene and protein interaction net-works, and ultimately to model these networks. The results of such integrative studies have several key advantages (re-viewed in [1-3]). In particular, using multiple sources of in-formation may allow us to reduce systematic noise inher-ently present in all types of experimental data. Integrative approaches are also important for the studies of complex diseases, such as cancer [4], since predicting the status of disease cases based on multiple biomarkers represents a starting point towards translating genomics research into clinical medicine. The integrative approach can also be used for predicting properties of one type of data based on other types of ‘omic’ (genomic, proteomic, etc) data [2, 5-7], for evaluating ‘omic’ datasets [8], and for functional prediction and inference [9-11]. Such a promise of the integrative ap-proach is based on the general assumption that, within a given genome, there exist inter-relationships between het-erogeneous types of genomic data [12]. Since even seem-ingly different data types describe various functional aspects of the same genome (e.g., the human genome), it seems rea-sonable to anticipate the existence of non-random associa-tions among them. However, the existence of such associa-tions needs to be verified and their strength needs to be

*Address correspondence to this author at the Gen*NY*sis Center for Ex-

cellence in Cancer Genomics, Department of Epidemiology and Biostatis-

tics, University at Albany, SUNY, One Discovery Drive, Rensselaer, NY

12144, USA; E-mail: [email protected]

quantified for each particular combination of data types [3]. In this manuscript, we use the term ‘association’ instead of ‘correlation’ in order not to confuse it with correlation be-tween expression profiles.

A number of studies have demonstrated the existence of non-random pairwise associations between different types of large-scale ‘omic’ datasets. ‘Non-random association’ or just ‘association’ in this context means that genes that are func-tionally related with respect to one data type also tend to be related with respect to another data type. For the first time, such an association was demonstrated on the example of the yeast interactome and transcriptome. Since interacting pro-teins must be present within the cell at the same time, genes that encode them should also be expressed during the same time intervals. Consistent with this reasoning, it was shown that yeast genes with similar expression profiles are more likely to encode interacting proteins than randomly chosen genes [13]. A related study of the yeast genome showed that genes encoding interacting proteins exhibit higher than aver-age co-expression [12]. This study also showed that the yeast protein-protein interaction (PPI) dataset contains a larger proportion of strongly co-expressed proteins, compared to their baseline proportion in the entire yeast proteome. Simi-larly, yeast proteins from the same protein complex show a stronger co-expression than random proteins [14]. The inter-actome-transcriptome correlation demonstrated in yeast was also demonstrated for multicellular organism C. elegans [10, 15-16]. Another important type of association is that be-tween expression and transcription factors (TF). It was shown for the yeast genome that when the same TFs target the same genes, these genes exhibit stronger co-expression than randomly selected ones [17].

The existence of two associations, PPI-expression and expression-TF locations, implies that there should also exist an association between PPI and TF locations. Consistent with this expectation, it was shown for proteins from the human N-methyl D-aspartate (NMDA) receptor that regula-tory regions of the genes that encode interacting proteins are

2 The Open Applied Informatics Journal, 2007, Volume 1 Hwang and Kuznetsov

targeted by similar sets of TFs [18-19]. The correlation be-tween PPI and TF data was also employed in order to dis-cover cooperative TF pairs that synergistically influence the expression of proteins that are located close to each other in the yeast protein-protein interaction network [20]. Correla-tions that involve biological pathways were also studied. Since genes that belong to the same pathway are functionally related, they can be expected to be co-expressed and co-regulated. An association between pathways and expression was shown for both tumor [21-22] and normal cells [22] from the human genome. An association between pathway data and data on transcriptional regulation was also demon-strated for several selected human pathways [18]. In yeast, relationships in a combination of three or more heterogene-ous types of genome-wide datasets have also been studied [23-25].

Most integrative studies have been done on the example of the yeast genome. Because of its relative simplicity, yeast is the best experimentally characterized eukaryotic organism for which many experimental large-scale datasets, such as PPI and locations of transcription factor binding sites (TFBS), are readily available. The human genome, on the other hand, is much more complex in nature and signifi-cantly harder to study experimentally. For instance, no com-prehensive experimental datasets on protein-protein interac-tions and TFBS locations are yet available for the human genome. Due to the absence of such experimental datasets, information about multiple genome-wide associations that involve PPI and TFBS locations in the human genome is lacking. A possible way to overcome this limitation is to study associations using computationally inferred genome-wide datasets.

In this work, we use a novel computational approach to perform a comprehensive analysis of four types of data that describe the following functional features of the human ge-nome: functional pathways, expression profiles, inferred protein-protein interactions, and inferred locations of tran-scription factor binding sites. We use inferred protein-protein interactions from OPHID (Online Predicted Human Interac-tome Database), the largest publicly available PPI database [26], that includes 8,687 human proteins. This PPI dataset is more than two orders of magnitude larger than the dataset of only 76 proteins used in a previously reported study of corre-lations involving the human interactome [18]. We analyze types of associations that have not been studied previously for the human genome, including associations between ex-pression and TFBS locations, PPI and expression, and path-way information and PPI. We study associations not only in pairwise combinations, but also in combinations of three and four data types.

METHODS

Sources of Genomic Data

This work deals with multiple heterogeneous sources of genomic data. We therefore need to use consistent unique gene identifiers for each of these sources. We utilize the hu-man genome annotation version 38 from the Ensembl data-base [27] to assign a unique id to each gene and keep this id for each data type. We obtained the following four types of data for the human genome using publicly available sources:

1. Biological pathways from the KEGG database [28]. In KEGG, each gene from the human genome is as-signed to one or more functional pathways. By map-ping KEGG identifiers onto Ensembl identifiers, we generated a list of 4,024 genes for which pathway an-notation is available.

2. Protein-protein interactions (PPI) from the OPHID database [26]. OPHID catalogs human protein-protein interactions that are either determined experimentally or inferred from known protein-protein interactions in model organisms (S. cerevisiae, C. elegans, D. mela-nogaster, and M. musculus). By mapping OPHID identifiers onto Ensembl identifiers, we generated a list of 8,687 genes whose protein products are anno-tated in the OPHID database.

3. Gene expression data from the SymAtlas database [29]. SymAtlas reports genome-scale gene expression measurements for 73 normal human tissues and 6 dis-ease state tissues hybridized to Affymetrix HG-U133A array. Two replicates were used for each tis-sue. In our analysis, we excluded disease state tissues and only used 73 normal tissues. Expression levels for each tissue were averaged over the two replicates. Thus, each gene was represented by an expression profile that consists of 73 data points. By mapping Affymetrix identifiers onto Ensembl identifiers, we generated a list of 12,306 genes whose expression profiles are annotated in the SymAtlas database.

4. The data on transcription factor binding sites (TFBS) were obtained as follows. First, we used the Ensembl human genome assembly version 38 [27] to retrieve regulatory upstream region of each gene. We define regulatory upstream region as a 2KB region upstream of the transcription start site. In Ensembl, a gene can be annotated as producing multiple transcripts, 1.3 transcripts per gene on average [30]. In cases when more than one transcript is annotated for a given gene, we use known transcript with most 5’ transcrip-tion start site. We choose known transcripts over novel transcripts because the former have more sup-porting evidence that the latter [30]. We used this procedure to retrieve regulatory upstream regions of all protein-coding genes (a total of 23,326 genes). Second, we used the Match software program [31] to scan the upstream regions for TFBS annotated in the TRANSFAC database [32]. The TRANSFAC data-base is a library of experimentally identified tran-scription factor binding sites represented in the form of a position weight matrix (PWM). Match is a tool that searches for putative TFBS in input DNA se-quences by using a library of PWMs. Match was run using the library of high-quality vertebrate PWMs and the option to minimize the number of false posi-tives. By parsing Match output, we obtained a list of putative TFBS found in the upstream regions of 23,326 human genes.

Conversion of Genomic Data into a Unified Matrix For-mat

Each type of genomic data was converted into a unified matrix format. In this format, a symmetric n by n matrix nu-

Bioinformatics Study of Functional Associations The Open Applied Informatics Journal, 2007, Volume 1 3

merically summarizes a particular type of functional rela-tionships observed among n genes. Each of the four types of data described above was converted into a matrix format as follows.

1. KEGG pathways are represented by matrix K (size 4,024 x 4,024). An element kij in K matrix is equal to 1 if products of genes i and j belong to at least one common KEGG pathway and 0 otherwise.

2. Protein-protein interactions are represented by matrix P (size 8,687 x 8,687). An element pij in P matrix has a binary value of 1 or 0, indicating the presence or ab-sence of protein-protein interaction between products of genes i and j.

3. Expression profiles are represented by matrix R (size 12,306 x 12,306). An element rij in R matrix is the Pearson correlation coefficient (PCC) between ex-pression profiles of genes i and j. For the cases when at least one gene in a pair (i,j) is mapped onto multi-ple Affymetrix probe sets (3,837 out of 12,306 genes), we calculate PCC between all probe set pairs that correspond to (i,j) and choose a PCC with the largest magnitude. Negative correlations in R matrix were set to zero. For analyses that involve computing association scores (see below), we use a binary ver-sion of R matrix in which all elements that have val-ues equal to or greater than 0.7 (strong correlation) are set to 1 and all elements that have values below 0.7 are set to 0.

4. The cis-similarity between promoter regions of genes is represented by matrix T (size 23,326 x 23,326). An element tij in T matrix is the number of unique TFBSs observed in the promoter regions of both gene i and j. Unique means that all occurrences of binding sites for the same TF are counted only once for each promoter region. For instance, if the promoter region of gene i contains 4 sites for transcription factor A and 1 site for transcription factor B, whereas the promoter re-gion of gene j contains 2 sites for transcription factor A and 3 sites for transcription factor B, the value of tij will be equal to 2. The idea of this definition of cis-similarity is to attempt to account for the number of common transcription factors that control both gene i and j.

When we study a combination of two or more types of data, we only use genes for which all types of required anno-tation are available and exclude genes with missing annota-tion. For example, when we study associations between K and P matrices, we take a set of genes for which both KEGG pathway and protein-protein interaction data are annotated.

Testing the Statistical Significance of the Associations Among Multiple Data Types

The main idea of presenting a particular type of genomic data as a symmetric matrix that describes a certain type of functional relationship between gene pairs is to reveal statis-tically significant functional associations among multiple matrices by using multiplication of equivalent matrix ele-ments. In general, when elements from k matrices of dimen-sion n, M1…Mk, that represent k types of genomic data for n genes are multiplied, and a final matrix is obtained,

F[i,j]=M1[i,j]*…*Mk[i,j] (note that this is an element-wise multiplication, not a conventional matrix product). In this final matrix F, gene pairs that exhibit strong associations across all k types of data will correspond to elements with large absolute value. The overall strength of functional asso-ciations within a group of n genes represented by k matrices can be quantified by computing the sum of all elements in the final matrix, S(n,k), as follows:

<

=

n

ji

k jiMjiMjiMknS ],[...],[],[),( 21 Eq. 1

If S(n,k) is significantly higher than that expected by chance, it will indicate that genes in the multiplied matrices exhibit a strong non-random association across k types of genomic data. We estimate the statistical significance of S(n,k) by comparing it to the distribution of random scores. A random score is obtained by randomly permuting elements in each matrix M1,…,Mk and then using these permuted ma-trices to obtain a score according to Eq. 1. For each matrix combination we generate 10,000 random scores. The p-value of the observed score, P(R(n,k) S(n,k)), is computed as fol-lows:

000,10

)),(),(()),(),((

knSknRNknSknRP = Eq. 2

where R(n,k) is random score and N(R(n,k) S(n,k)) is the number of random scores that are equal to or larger than S(n,k). We applied the Shapiro-Wilk normality test and found random association scores to be normally distributed (data not shown). Histograms of the distributions of random scores can be found in Supplementary information. Since most p-values obtained from random simulations are zero, we use the z-score to rank the associations:

),(

),(),()),((

knR

knRknSknSscorez

><= Eq. 3

where <R(n,k)> is the average and R(n,k) is the standard de-viation of the random score.

RESULTS

Representation of Genomic Data in a Matrix Format

For each gene in the human genome, we collected four types of raw functional data (see Methods for details): KEGG pathways, protein-protein interactions, expression profiles, and putative transcription factor binding sites (see Table 1). Each raw data type was converted into a unified matrix format, abbreviated as follows: K (KEGG pathway matrix), P (PPI matrix), R (co-expression matrix), and T (cis-similarity matrix). In this format, each symmetric n by n ma-trix numerically summarizes a particular type of functional relationships observed among n genes. An element kij in K matrix has a binary value of 1 if products of genes i and j belong to at least one common KEGG pathway and 0 other-wise. An element pij in P matrix has a binary value of 1 or 0, indicating the presence or absence of protein-protein interac-tion between products of genes i and j. An element rij in R matrix is the Pearson correlation coefficient (PCC) between expression profiles of genes i and j. An element tij in T ma-trix is a degree of cis-similarity between genes i and j. We define cis-similarity as the number of unique TFBSs shared by the promoter regions of genes i and j. Unique means that


all occurrences of TFBSs for the same transcription factor are counted only once for each promoter region. Thus, our definition of cis-similarity is the number of common TFs involved in transcriptional control of both gene i and j.

Associations Observed Among Multiple Types Of Ge-

nomic Data

In this section, we perform a qualitative study of ge-nome-wide associations observed among the four types of genomic data. The idea of this study is to examine whether the properties of genes with respect to one type of functional data are correlated with other types of functional data. For example, we can classify pairs of genes into interacting and non-interacting categories and examine the average correla-tion coefficient between their expression profiles in order to see whether expression profiles of genes whose products interact tend to have a higher correlation coefficient than the profiles of non-interacting ones. Here, we study global ge-nome-wide associations for the following combinations of data types: K-P, R-T, K-R, P-T, K-P-R, and K-P-T and dem-onstrate the existence of potentially significant relationships observed among these data types. A rigorous statistical analysis of the significance of associations for all possible combinations of data types that confirms the qualitative trends discussed here is presented in the following sections of the manuscript.

First, we analyze the associations between functional pathways and protein-protein interactions (K-P association). The comparison of pathway information for interacting and non-interacting proteins shows that 40.82% of interacting protein pairs share at least one functional pathway (meaning that both proteins in the pair belong to the same pathway), whereas only 5.55% of non-interacting protein pairs share pathway annotation (Fig. 1A). This means that interacting protein pairs are seven times more likely to participate in the same pathway than non-interacting protein pairs. Analysis of the reverse relation shows that if two proteins participate in the same pathway, they are eleven times more likely to inter-act than proteins from different pathways (Fig. 1B).

Second, we analyze the associations between co-expression and cis-similarity of promoter regions (R-T asso-ciation). This analysis shows that, on average, correlation between expression profiles of genes that share common TFBS is higher (PCC=0.375) than that between expression profiles of genes that do not share any common TFBS (PCC=0.369) (Fig. 1C). Analysis of the reverse relation shows that an increase in the level of co-expression of gene pairs is associated with an increase in the number of com-mon TFBS found in their promoter regions (Fig. 1D). These results confirm to an empirical expectation that co-expressed

genes should have similar cis-profiles and vice versa. How-ever, the trends shown in Figs. (1C) and (1D) are very subtle and their statistical significance is not obvious. One possible reason of weak trends is that the computational identification of putative TFBS via sequence motif-based methods is in-herently prone to noise because of a very high percentage of false positive predictions [33]. We will address the issue of statistical significance in the next section of the manuscript.

Third, we analyze the following three types of associa-tions: K-R, P-R, and K-P-R. We divided all gene pairs into six categories according to whether their products are inter-acting and/or participating in same functional pathways and compared the average correlations between expression pro-files for these six categories (Fig. 1E). From right to left in Fig. (1E), the largest average PCC between expression pro-files is found for gene pairs that both interact and participate in same pathways (PCC = 0.4331), whereas the smallest av-erage PCC is found for gene pairs that neither interact nor participate in same pathways (PCC = 0.3216). We also ob-serve that the average PCC is higher for gene pairs that par-ticipate in same pathways (PCC = 0.4028) than for interact-ing pairs (PCC = 0.3773). These observations suggest that, with respect to concerted expression, genes from the same pathway act as a more cohesive biological module than genes producing physically interacting proteins. Experimen-tal evidence shows that interacting proteins from the same complex are not necessarily produced by co-regulated genes. For example, cyclin-dependent kinase and cyclin together form a protein complex. While the former is produced from a constantly transcribed gene, the latter is produced in a regulated manner [13].

Fourth, we analyze K-T, P-T, and K-P-T associations by computing the average number of common TFBS for the same six categories of gene pairs described above. The re-sults of this analysis, shown in Fig. (1F), reveal a trend very similar to the one shown in Fig. (1E): the largest number of common TFBS is observed for gene pairs that both interact and participate in the same pathways, whereas the smallest number of common TFBS is observed for gene pairs that neither interact nor participate in the same pathway. These two related trends indirectly indicate that the level of co-expression (measured by PCC) and the cis-similarity (meas-ured by the number of common TFBS) are correlated with each other, which is in agreement with the direct relationship between them shown in Figs. (1C) and (1D). The small dif-ferences in the number of common TFBS observed in Fig. (1F) can be attributed to the fact that the computational pro-cedure for the identification of putative TFBS produces a very large number of false positives [33].

Table 1. Four Types of Data Used in this Study

Data Type Description Source Genes

K Functional pathways KEGG [28] 4,024

P Protein-protein interactions OPHID [26] 8,687

R Expression profiles SymAtlas [29] 12,306

T Putative TFBS found in the promoter regions Ensembl [27], TRANSFAC [32] 23,326


Statistical Significance of the Associations Among Multi-ple Types of Genomic Data

The qualitative analyses shown in the previous section indicate the existence of potentially significant associations among various types of ‘omic’ data. In this section, we use a rigorous quantitative approach to evaluate the statistical sig-nificance of the observed associations on the genome-wide scale. Given two or more matrices that represent particular

types of ‘omic’ data for a group of genes, we measure the strength of association among these data types by means of an association score. This association score is defined as the sum of products between corresponding elements of the ma-trices under consideration (Eq. 1, see Methods for details). Statistical significance of the observed association score is estimated by comparing it to the distribution of random as-sociation scores obtained from randomly permuted matrices.

Fig. (1). Associations exist among the four types of functional data. (A) Interacting protein pairs are more likely to participate in the same

pathway than non-interacting protein pairs. (B) Protein pairs from the same pathway are more likely to interact than protein pairs from dif-

ferent pathways. (C) A pair of genes that share common TFBS in the promoter regions shows a higher correlation between expression pro-

files than a pair without any shared TFBS. (D) An increase in correlation between expression profiles is associated with an increase in the

number of shared TFBS. (E) Pairs of proteins from the same pathway and/or pairs of interacting proteins are more likely to show correlated

expression. (F) Pairs of protein from the same pathway and/or pairs of interacting proteins are more likely to share common TFBS in their

promoter regions.


Since we have four matrices that correspond to the four types of data, there are eleven possible combinations of two, three and four matrices. P-values and z-scores for each combina-tion are shown in Table 2. Histograms of all random distri-butions can be found in the Supplementary information, with three selected distributions shown in Figs. (2-4).

The results shown in Table 2 indicate that all eleven combinations of data types demonstrate significant associa-tions as evidenced by low p-values. Below, we briefly dis-cuss biological implications of each association. The results for pairwise combinations indicate:

K-P (z-score=160.97, p=0) - the existence of a highly significant association between protein-protein interactions and protein function. Since in our methodology associations are not directional, K-P association is equivalent to P-K as-sociation, thus implying that interacting proteins tend to par-ticipate in the same functional pathway, and vice versa, pro-teins from the same functional pathway tend to interact.

Fig. (2). The distribution of random K-P scores.

K-R (z-score=85.35, p=0) - the existence of a highly sig-nificant association between co-expression and gene func-tion. It shows that genes participating in the same pathway

tend to be co-expressed, and vice versa, co-expressed genes tend to participate in the same pathway.

R-T (z-score=75.63, p=0) - co-expressed genes tend to share similar cis-profiles, and vice versa, genes with similar cis-profiles tend to be co-expressed.

P-R (z-score=7.70, p=0) - genes that encode interacting proteins tend to be co-expressed, and vice versa, co-expressed genes tend to encode interacting proteins.

K-T (z-score=7.59, p=0) - genes from the same pathway tend to have similar cis-profiles, and vice versa, genes with similar cis-profiles tend to participate in the same pathway.

Fig. (3). The distribution of random K-R scores.

P-T (z-score=1.66, p=0.0498) - the existence of a mar-ginally significant association between protein-protein inter-actions and the similarity of cis-profiles of the genes that encode interacting proteins.

The results for all combinations of three data types, de-scribed below, also demonstrate highly statistically signifi-cant genome-wide associations:

K-P-T (z-score=142.26, p=0) - the existence of a highly significant association that links gene function, K, interac-

Table 2. P-Values and Z-Scores Estimated from the Random Permutation Experiment

Type of Association Genes P-Value Z-Score

K-P 2,424 0 160.97

K-R 3,271 0 85.35

R-T 11,775 0 75.63

P-R 6,784 0 7.70

K-T 3,939 0 7.59

Pairwise combinations

P-T 8,513 0.0498 1.66

K-P-T 2,424 0 142.26

K-P-R 2,154 0 82.22

K-R-T 3,271 0 73.90 Triple combinations

P-R-T 6,784 0 7.47

Quadruple combination K-P-R-T 2,154 0 75.65


tions between gene products, P, and cis-similarity of the promoter regions, T. This association implies that genes from the same pathway both tend to code for interacting pro-teins and share a similar set of TFs in their promoter regions.

Fig. (4). The distribution of random K-P-R scores.

K-P-R (z-score=82.22, p=0) - the existence of a highly significant association that links gene function, K, interac-tions between gene products, P, and co-expression, R. Bio-logically, this association is similar to K-P-T association and implies that genes from the same pathway both tend to code for interacting proteins and to be co-expressed.

K-R-T (z-score=73.90, p=0) - the existence of a highly significant association that links gene function, K, co-expression, R, and cis-similarity of the promoter regions, T. This association implies that genes from the same pathway tend to be both co-expressed and have a similar set of TFs in their promoter regions.

P-R-T (z-score=7.47, p=0) - the existence of a significant association that links interactions between gene products, P, co-expression, R, and cis-similarity of the promoter regions, T. This association implies that genes whose products inter-act tend to be co-expressed and have a similar set of TFs in their promoter regions. However, it should be pointed out that the strength of P-R-T association is much weaker than that of other triple associations as indicated by a considera-bly lower z-score.

Finally, the results for the combination of all four data types, K-P-R-T, indicate that this quadruple association is also highly significant (z-score=75.65, p=0). This association indicates that genes from the same pathway simultaneously tend to encode interacting proteins, be co-expressed, and have a similar set of TFs in their promoter regions.

Pathway-Level Analysis of the Associations Among Data Types

The analysis reported in the previous section summarizes global genome-wide relations among data types by consider-ing all genes in the human genome simultaneously. A similar analysis can be performed by considering a group of genes that belong to a particular functional category. A good ex-ample of a functional category is a functional pathway, which can be considered as a biological module that carries out a specific genomic function. Depending on the function

of the pathway, one may expect certain pathway-specific associations to be more pronounced than the others. In this section, we use the classification of functional pathways from the KEGG database [28]. The main difference from the global analysis reported in the previous section is that path-way-level analysis is done for a group of functionally related genes that belong to a particular KEGG pathway. This kind of analysis enables us to categorize the associations between the pathway data and other types of genomic data and to determine which types of associations are most profound in a particular functional category. Since the TFBS data seem to be noisy and therefore least reliable, we use only the PPI data and the gene expression data for the pathway-level analysis. This leaves us with three combinations to analyze for pathway-level associations: K-P, K-R and K-P-R. These three combinations provide the following biological informa-tion for a given pathway: K-P describes its relative enrich-ment in interacting proteins, K-R describes its relative en-richment in co-expressed genes, and K-P-R describes its relative enrichment in genes that are both co-expressed and code for interacting proteins. When we use Eqs. 2,3 to ana-lyze a group of m genes that belong to a particular pathway, each random m by m matrix for a given type of data is ob-tained by randomly sampling, without replacement, m genes from a list of all human genes annotated with this particular data. Pathways containing less than five annotated genes were excluded from this analysis.

Out of 174 pathways annotated in the KEGG database for the human genome, we identified 98 pathways that are sig-nificantly (p<0.05) enriched in interacting proteins, 34 path-ways that are significantly enriched in co-expressed genes, and 75 pathways that are significantly enriched in genes that are both co-expressed and code for interacting proteins. Lists for all pathways and all combinations, ranked by z-score, are given in Supplementary information. Top 10 scoring path-ways for each combination are shown in Tables 3, 4 and 5. It should be noted that when pathways are analyzed with

Table 3. Top 10 Pathways Enriched in Interacting Proteins

KEGG ID Z-Score Pathway Name

hsa03050 237.19 Proteasome

hsa03020 134.22 RNA polymerase

hsa03010 86.94 Ribosome

hsa00193 61.66 ATP synthesis

hsa00240 48.63 Pyrimidine metabolism

hsa04110 44.15 Cell cycle

hsa03022 40.98 Basal transcription factors

hsa04130 39.30 SNARE interactions in vesicular transport

hsa00020 39.19 Citrate cycle (Tca cycle)

hsa04350 34.76 Tgf-beta signaling pathway

respect to concomitant enrichment in co-expressed genes whose protein products also interact (the triple K-P-R asso-ciation), several additional pathways emerge as significant (Table 5 and Supplementary Table 3). For example, two


pathways (‘Glutamate metabolism’ and ‘Glutathione me-tabolism’) are identified as showing significant concomitant enrichment, even though they do not show enrichment in interacting proteins or co-expressed genes. Some pathways are concomitantly enriched even though they show enrich-ment in either interacting proteins or co-expressed genes, but not both. For example, ‘Cholera’ pathway shows very strong concomitant enrichment (top 5

th in Table 5), but it does not

show enrichment in co-expressed genes. Similarly, ‘Olfac-tory transduction’ pathway shows a significant concomitant enrichment without being enriched in interacting proteins. These observations indicate that combining multiple types of genomic data reveals additional functional features of indi-vidual pathways that cannot be revealed by studying simple pairwise associations.

Table 4. Top 10 Pathways Enriched in Co-Expressed Genes


hsa04080 9.73 Neuroactive ligand-receptor interaction

hsa04630 5.11 Jak-Stat signaling pathway

hsa04620 4.94 Toll-like receptor signaling pathway

hsa00190 4.91 Oxidative phosphorylation

hsa04020 4.69 Calcium signaling pathway

hsa00602 4.60 Glycosphingolipid biosynthesis-neo-lactoseries

hsa04010 3.86 MAPK signaling pathway

hsa04060 3.81 Cytokine-cytokine receptor interaction

hsa04730 3.57 Long-term depression

hsa04664 3.51 Fc epsilon RI signaling pathway

Table 5. Top 10 Pathways Enriched in Both Interacting and

Co-Expressed Genes


hsa00193 78.58 ATP synthesis

hsa03050 76.25 Proteasome

hsa03020 44.87 RNA polymerase

hsa00190 36.83 Oxidative phosphorylation

hsa05110 30.64 Cholera

hsa04350 29.41 TGF-beta signaling pathway

hsa04660 27.01 T cell receptor signaling pathway

hsa03010 25.52 Ribosome

hsa05120 25.18 Epithelial cell signaling in helicobacter pylori infection

hsa04610 23.94 Complement and coagulation cascades

DISCUSSION

In general, the results of the quantitative analysis of the genome-wide pairwise associations are consistent with the qualitative study performed on the same datasets (see Fig. 1) and discussed in the first section of Results, thus confirming

the utility of the proposed approach. For instance, Fig. (1E) shows that the average correlation between expression pro-files is larger for gene pairs from the same pathway, K, (PCC=0.4028) than for gene pairs that encode interacting proteins (PCC=0.3773). This observation is consistent with a larger z-score observed for K-R association (z-score=85.35) compared to that for P-R association (z-score=7.70). Simi-larly, Fig. (1F) shows that the average number of shared TFBS is larger for gene pairs from the same pathway (3.2869) than for gene pairs whose products interact (3.2705). This observation is also consistent with a larger z-score observed for K-T association (z-score=7.59) compared to that for P-T association (z-score=1.66). If we assume that K and P matrices contain similar amounts of noise, then the observation that the z-scores for K-R and K-T associations are larger than those for P-R and P-T suggests that transcrip-tional co-regulation is more important for genes from the same pathway than for genes that encode interacting pro-teins. It should also be noted that, to the best of our knowl-edge, out of the six pairwise combinations of data types util-ized in this work, three (P-R, R-T, and K-P) have never been studied for the human genome.

The strongest pairwise associations, indicated by very high z-scores, are observed for combinations involving pathway data, K-P and K-R. This observation is consistent with empirical expectations and confirms that genes from the same functional pathway tend to be co-expressed and code for interacting proteins. The only marginally significant ge-nome-wide association is observed between PPI data and cis-similarity of promoter regions (P-T combination, p=0.0498, Table 2). The relatively low z-scores for two associations involving the T matrix (P-T and K-T) are not straightforward to interpret. On one hand, a large amount of noise present in matrix T may dampen real biological associations. On the other hand, the R-T association is quite significant (z-score=75.63) despite the noise present in the T matrix. The observation that P-R association, which is related to P-T, also has a relatively low z-score of 7.7 provides an additional argument in favor of the assumption that the weakness of the genome-wide P-T association may reflect a real biological phenomenon.

The application of our methodology to study associations in groups of genes from individual functional pathways shows that pathways enriched in interacting proteins (K-P association, Table 3 and Supplementary Table 1) are mostly the ones for genetic information processing. These pathways tend to contain large protein complexes, such as the ribo-some and DNA/RNA polymerases. Pathways enriched in co-expressed genes (K-R association, Table 4 and Supplemen-tary Table 2) are mostly the pathways for environmental information processing. These pathways can be thought of as biological modules whose genes need to be expressed in a concerted manner in response to external stimuli. Metabolic pathways seem to be under-represented in the list of path-ways enriched in co-expressed genes. There are 112 meta-bolic pathways, comprising 64% of all 174 annotated path-ways. However, out of the total of 34 pathways significantly enriched in co-expressed genes, only six (18%) are metabolic pathways. This observation is consistent with previously reported results that metabolic pathways do not show similar cis-profiles [18].


Another possible application of the proposed methodol-ogy is to benchmark the quality of various large-scale datasets. In this work, we used PPI from the OPHID data-base [26], where about 60% of all annotated interactions were inferred computationally, rather than obtained experi-mentally. Obviously, the quality of this inference needs to be validated. Since proteins that participate in the same func-tional pathway often form multi-protein complexes and can be expected to interact, the strength of K-P association can be used as an indicator of the non-randomness of PPI annota-tions. The fact that, according to our results, K-P association ranks highest among all pairwise combinations studied, sug-gests that the assignment of PPI in groups of functionally related proteins is highly non-random, thus confirming the quality of the OPHID annotation. Therefore, the present work can also be considered as an independent validation of OPHID, in addition to the validation provided by the authors of this database. A similar approach can be used to bench-mark other types of data. For instance, given several meth-ods for finding TFBS in promoter regions, the R-T associa-tion experiment can be used as a quantitative evaluation pro-cedure to benchmark which of these methods gives the best correlation with expression data.

CONCLUSIONS

We presented a methodology for quantifying the signifi-cance of associations observed among multiple heterogene-ous types of ‘omic’ data. We used this methodology to ana-lyze and rank associations among four types of data in hu-mans: functional pathways, protein-protein interactions, ex-pression profiles, and transcription factor binding sites. Us-ing the largest datasets currently available for the human genome, we showed that associations in all combinations of data types are significantly higher than random expectation. We find that pathways involved in genetic information proc-essing are enriched in interacting proteins, and that signaling pathways are enriched in co-expressed genes. We show that combinations of multiple data types provide additional in-sights into the properties of functional pathways. This work can also be considered as an independent validation of the inferred protein-protein interactions annotated in the OPHID database.

ABBREVIATIONS

PPI = protein-protein interaction

TF = Transcription factor

TFBS = Transcription factor binding site

PCC = Pearson correlation coefficient

K = KEGG pathway data matrix

P = PPI data matrix

R = Co-expression data matrix

T = Cis-similarity data matrix

REFERENCES

[1] M. Gerstein, N. Lan, R. Jansen, “Proteomics. Integrating interac-tomes,” Science, vol. 295, pp. 284-287, January 2002.

[2] D. Greenbaum, N. M. Luscombe, R. Jansen, J. Qian, M. Gerstein, “Interrelating different types of genomic data, from proteome to se-

cretome: 'oming in on function,” Genome Res., Vol. 11, pp. 1463-1468, September 2001.

[3] M. Vida, “A biological atlas of functional maps,” Cell, Vol. 104,

pp. 333-339, February 2001. [4] S. Wachi, K. Yoneda, R. Wu, “Interactome-transcriptome analysis

reveals the high centrality of genes differentially expressed in lung cancer tissues,” Bioinformatics, Vol. 21, pp. 4205-4208, February

2005. [5] A. Drawid, M. Gerstein, “A Byesian system integrating expression

data with sequence patterns for localizing proteins: comprehensive application to the yeast genome,” J. Mol. Biol., Vol. 301, pp. 1059-

1075, December 2001. [6] J. Qian, J. Lin, N. M. Luscombe, H. Yu, M. Gerstein, “Prediction

of regulatory networks: genome-wide identification of transcription factor targets from gene expression data,” Bioinformatics, Vol. 19,

pp. 1917-1926, August 2003. [7] L. V. Zhang, S. L. Wong, O. D. King, F. P. Roth, “Predicting co-

complexed protein pairs using genomic and proteomic data integra-tion,” BMC Bioinformatics, Vol. 5, pp. 38, April 2004.

[8] J. S. Bader, A. Chaudhuri, J. M. Rothberg, J. Chant, “Gaining confidence in high-throughput protein interaction networks,” Nat.

Biotechnol. Vol. 22, pp. 78-85, January 2004. [9] C. S. Goh, T. A. Gianoulis, Y. Liu, et al. “Integration of curated

databases to identify genotype-phenotype associations,” BMC Ge-nomics, Vol. 7, pp. 257, October 2006.

[10] K. C. Gunsalus, H. Ge, A. J. Schetter, et al. “Predictive models of molecular machines involved in Caenorhabditis elegans early em-

bryogenesis,” Nature, Vol. 436, pp. 861-865, August 2005. [11] I. Lee, S. V. Date, A. T. Adai, E. M. Marcotte, “A probabilistic

functional network of yeast genes,” Science, Vol. 306, pp. 1555-1558, November 2004.

[12] A. Grigoriev, “A relationship between gene expression and protein interactions on the proteome scale: analysis of the bacteriophage

T7 and the yeast Saccharomyces cerevisiae,” Nucleic Acids Res., Vol. 29, pp. 3513-3519, September 2001.

[13] H. Ge, Z. Liu, G. M. Church, M. Vidal, “Correlation between tran-scriptome and interactome mapping data from Saccharomyces cer-

evisiae,” Nature Genet., Vol. 29, pp. 482-486, December 2001. [14] R. Jansen, D. Greenbaum, M. Gerstein, “Relating whole-genome

expression data with protein-protein interactions,” Genome Res.Vol. 12, pp. 37-46, January 2002.

[15] A. J. Walhout, J. Reboul, O. Shtanko, et al. “Integrating interac-tome, phenome, and transcriptome mapping data for the C. elegans

germline,” Curr. Biol., Vol. 12, pp. 1952-1958, November 2002. [16] S. Li, H. Ge, T. Hao, et al. “A map of the interactome network of

the metazoan C. elegans,” Science, Vol. 303, pp. 540-543, January 2004.

[17] H. Yu, N. M. Luscombe, J. Qian, M. Gerstein, “Genomic analysis of gene expression relationships in transcriptional regulatory net-

works,” Trends Genet., Vol. 19, pp. 422-427, August 2003. [18] S. Hannenhalli, S. Levy, “Transcriptional regulation of protein

complexes and biological pathways,” Mamm. Genome, Vol. 14, pp. 611-619, September 2003.

[19] O. Alter, G. H. Golub, “Integrative analysis of genome-scale data by using pseudoinverse projection predicts novel correlation be-

tween DNA replication and RNA transcription,” Proc. Natl. Acad. Sci. USA, Vol. 101, pp. 16577-16582, November 2004.

[20] N. Nagamine, Y. Kawada, Y. Sakakibara, “Identifying cooperative transcriptional regulations using protein-protein interactions,” Nu-

cleic Acids Res., Vol. 33, pp. 4828-4837, August 2005. [21] H. H. Yang, Y. Hu, K. H. Buetow, M. P. Lee, “A computational

approach to measuring coherence of gene expression in pathways,” Genomics, Vol. 84, pp. 211-217, July 2004.

[22] R. Huang, A. Wallqvist, D. G. Covell, “Comprehensive analysis of pathway or functionally related gene expression in the National

Cancer Institute's anticancer screen,” Genomics, Vol. 87, pp. 315-328, March 2006.

[23] A. Tanay, R. Sharan, M. Kupiec, R. Shamir, “Revealing modularity and organization in the yeast molecular network by integrated

analysis of highly heterogeneous genomewide data,” Proc. Natl. Acad. Sci. USA, Vol. 101, 2981-2986, March 2004.

[24] D. Hwang, A. G. Rust, L. Hood, et al. “A data integration method-ology for systems biology,” Proc. Natl. Acad. Sci. USA, Vol. 102,

pp. 17296-17301, November 2005. [25] P. Carmona-Saez, M. Chagoyen, A. Rodriguez, O. Trelles, J. M.

Carazo, A. Pascual-Montano, “Integrated analysis of gene expres-sion by association rules discovery,” BMC Bioinformatics, Vol. 7,

pp. 54, February 2006.


[26] K. R. Brown, I. Jurisica, “Online predicted human interaction data-

base,” Bioinformatics, Vol. 21, pp. 2076-2082, May 2005. [27] T. Hubbard, D. Andrews, M. Caccamo, et al. “Ensembl 2005,”

Nucleic Acids Res., Vol. 33, pp. D447-D453, January 2005. [28] M. Kanehisa, S. Goto S, “KEGG: Kyoto Encyclopedia of Genes

and Genomes,” Nucleic Acids Res., Vol. 28, pp. 27-30, January 2000.

[29] A. I. Su, H. Lapp, J. Zhang et al. “A gene atlas of the mouse and human protein-encoding transcriptomes,” Proc. Natl. Acad. Sci.

USA, Vol. 101, pp. 6062-6267, April 2005. [30] V. Curwen, E. Eyras, T. D. Andrews, et al. “The Ensembl auto-

matic gene annotation system,” Genome Res., Vol. 14, pp. 942-950, May 2004.

[31] A. E. Kel, E. Gossling, I. Reuter, E. Cheremushkin, O. V. Kel-

Margoulis, E. Wingender, “MATCH: a tool for searching transcrip-tion factor binding sites in DNA sequences,” Nucleic Acids Res.,

Vol. 31, pp. 3576-3579, July 2003. [32] V. Matys, E. Fricke, R. Geffers, et al. “TRANSFAC: transcrip-

tional regulation, from patterns to profiles,” Nucleic Acids Res., Vol. 31, pp. 374-378, January 2003.

[33] M. Robinson, Y. Sun, R. T. Boekhorst, et al. “Improving computa-tional predictions of cis-regulatory binding sites,” Pac. Symp. Bio-

comput., Vol. 11, pp. 391-402, November 2006.

Received: May 15, 2007 Revised: May 22, 2007 Accepted: May 26, 2007

the open applied informatics journal, 1-10 1 ...the open applied informatics journal, 2007, 1, 1-10...

Documents